Google made its conversational video model generally available on August 27, 2026. The model id is gemini-omni-1.1-flash, and it replaces the gemini-omni-flash-preview endpoint that shipped in public preview on June 30. If you built anything against the preview, you have a deadline: Google turns the preview endpoint off on September 30, 2026.
The GA release is not a version bump with nothing behind it. Four capabilities landed with it, and each one closes a gap that made the preview hard to use in production: scene extension out to 40 seconds, first-and-last-frame control, a cheap 360p draft mode, and 4K upscaling. This post covers what changed, what it costs, where the model runs, and what it still refuses to do.
If you want the background on what the Omni family is and why Google built a reasoning-first video model in the first place, start with what is Gemini Omni. This post assumes you already know that part and want the 1.1 details.
What Gemini Omni 1.1 Flash does
Omni 1.1 Flash takes text, images, and video as input, and returns video. It runs on the Interactions API rather than generateContent, which is the same surface Google uses for stateful, multi-turn generation. That matters more than it sounds: you can generate a clip, then send a follow-up turn that edits the clip you just made, without re-uploading anything.
The five task shapes the model supports:
- Text to video: a prompt goes in, a clip comes out.
- Image to video: one image becomes the opening frame, or a style guide for motion.
- Reference to video: up to three reference clips of three seconds each map movement, look, and character consistency onto a new scene.
- Edit: change something inside a video you already generated, in conversation.
- Extend: append another 10 seconds to the end of a clip.
Audio on reference clips is ignored. The model reads them for motion and appearance only.
Scene extension now reads 10 seconds of context
This is the headline change. The preview model looked at the final second of a clip before continuing it, which is roughly enough context to keep the color palette stable and not much else. Characters drifted. Camera moves reset. You could extend a video, but you could not extend a shot.
Omni 1.1 analyzes up to 10 seconds of prior context, and Google describes the result as “improved visual consistency and narrative adherence”. You extend in 10-second increments up to a cumulative 40 seconds.
Two limits to plan around. Extension only appends to the end of a clip: you cannot prepend footage or splice into the middle. And uploaded videos max out at 10 seconds of input unless you stay inside a multi-turn interaction, where the model already holds the earlier state.
The 40-second scene extension walkthrough covers the request shape and where the seams show up.
Keyframe control between two images
You can now hand the model a first frame and a last frame, plus a prompt describing the movement between them, and it generates the footage that connects the two. Google points at camera orbits, zoom transitions, and looping clips as the obvious uses.
For anyone building a video tool rather than a toy, this is the feature that makes output predictable. Text-to-video is a slot machine. Frame interpolation gives you two fixed points and lets the model solve for the middle, which is a much smaller space to get wrong.
360p drafting, and why it changes the cost math
Omni 1.1 generates 360p previews up to 60% faster than 720p, at one third of the cost. Google is explicit about the intended workflow: draft at 360p until the prompt is right, then re-render the keeper at 720p, 1080p, or 4K.
That single change is worth more to a production budget than any other item in the release. Video generation is expensive per second, and most of what you generate during prompt iteration gets thrown away. Paying a third as much for the throwaway passes is the difference between exploring freely and rationing your attempts. The full pricing breakdown has the per-second math.
Resolution is set in the response format, and 1080p and 4k are upscales of the generated frames rather than native renders.
Where it runs
Omni 1.1 Flash is available through:
- The Gemini API in Google AI Studio, for developers
- The Gemini Enterprise Agent Platform API, for enterprise deployments
- Google Flow, for AI Plus, Pro, and Ultra subscribers
- The Gemini app, where subscribers get scene extension
Google also named launch partners running the model inside their own products: Adobe Firefly, Figma Weave, Runway, and GMI Cloud. If you use any of those tools, you are already sending traffic to this model.
There’s no free tier on the API. Unlike the text Flash models, where AI Studio gives you a rate-limited free lane, every second of Omni output bills from the first request. The model itself is free on YouTube Shorts and YouTube Create, which is a different product with different limits; the free-access rundown covers what each path gets you.
What it still can’t do
Read this list before you scope a feature around the model.
- No prompt controls you might expect. System instructions,
temperature,top_p, stop sequences, and negative prompts are all unsupported. If you want to exclude something, you write that into the regular prompt. - Regional restrictions. Uploading and editing videos is unavailable in the European Economic Area, Switzerland, and the UK, as is editing images that contain minors. Model-generated videos remain editable in those regions.
- Recognizable people. Editing certain identifiable individuals is blocked.
- Dialogue on uploaded video. You can extend an uploaded clip silently, or work in a multi-turn interaction, but you cannot add dialogue when extending someone else’s upload.
- One video at a time for reasoning. The model won’t reason across multiple input videos.
- English only, in practice. Google says English is fully supported and other languages are untested.
Every output carries a SynthID watermark, invisible to viewers and detectable programmatically. Plan for that if your product makes provenance claims.
Omni 1.1 Flash or Veo 3.1?
Google now ships two video generation families on the same API key, which is a genuinely confusing situation. The short version: Veo 3.1 is the cinematic renderer with native audio, and Omni 1.1 Flash is the conversational one you talk to across turns. Omni’s 720p second costs about a quarter of a standard Veo 3.1 second, and Veo has no equivalent to the multi-turn edit loop.
The side-by-side comparison goes through the cases where each one wins. If you already have a Veo integration, the Veo 3.1 API guide is still accurate; nothing in this release deprecates it.
Migrating off the preview endpoint
Anything pointing at gemini-omni-flash-preview stops working after September 30, 2026. The migration itself is usually a one-line change to the model string, but two things are worth checking before you assume that’s all:
- Resolution defaults. 720p is the default now, and the 360p and 4K options are new. If your code assumed a fixed output size, confirm what you’re actually getting back.
- Response size. Videos over 4MB come back as a URI instead of inline base64. If your handler only reads
output_video.data, a higher resolution will silently hand you nothing.
The cleanest way to catch both is to save the request in an API client and diff the responses between the two model ids before you cut over. Apidog handles this well: put the model id in an environment variable, keep one saved request, flip the environment, and compare. The API walkthrough sets that up step by step.
FAQ
What’s the model id for Gemini Omni 1.1 Flash? gemini-omni-1.1-flash. The retiring preview id is gemini-omni-flash-preview.
When did Gemini Omni 1.1 Flash come out? August 27, 2026, as the GA release. The preview shipped June 30, 2026.
Is Gemini Omni 1.1 Flash free? No. There’s no free tier for Omni, so every generation bills. Text models like Gemini 3.6 Flash still have a free AI Studio lane; this one doesn’t.
How long can a Gemini Omni video be? Clips generate at 10 seconds, and scene extension takes them to 40 seconds cumulative in 10-second steps.
Does Gemini Omni generate audio? The docs cover video output and ignore audio on reference clips. Veo 3.1 is the model with native audio generation. If sound matters, start there.
Can I edit a video I filmed myself? Yes, up to 10 seconds of input, outside the EEA, Switzerland, and the UK. Upload it with the Files API and pass the URI as input.
Omni 1.1 Flash is the first version of this model that’s shaped like something you’d ship. Scene extension that holds a shot together, keyframes that make output predictable, and a draft mode that makes iteration affordable. Wire up one saved request against the GA model id, check the response shape at every resolution you plan to use, and get off the preview endpoint before the end of September. Download Apidog if you want that check saved before the deadline rather than after it.
